Neural networks for assessing the risk of windthrow on the forest division level: a case study in southwest Germany

The paper describes a methodological approach for risk assessment on the forest division level especially in secondary coniferous forests based on regular management records using an artificial neural network (ANN). Database of the investigation are records of salvage cuttings mainly due to storm damage and of inventory data for a State Forest of the South Black Forest, Germany in the period from 1976 to 1985. The data on salvage cuttings were recorded on the forest division level, an area with a mean size of 30 ha that contains several forest stands. The database has partly to be considered as rather unsharp or fuzzy and makes classical statistical analysis very difficult. The paper provides a description of the methodological approach, especially of the structure of the neural network, a three layered feed forward network with backpropagation algorithm and the tangens hyperbolicus as activation function. First experiences with the application of the method were made by predicting the risk for storm damage within one decade for a state forest containing 139 divisions (around 4,000 ha). Two randomly selected training sets were presented to the network and produced a very satisfying learning effect. The validation for two test sets showed a good result for the lower damage class and a tendency for overestimating the damage in the higher damage class. A classical logistic regression model that was compared to the ANN systematically underestimated the damage for the training and test data sets.